bioRxiv preprint doi: https://doi.org/10.1101/148684; this version posted June 12, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. Comprehensive genome and transcriptome analysis reveals genetic basis for gene fusions in cancer Nuno A. Fonseca1*, Yao He2*, Liliana Greger1, PCAWG3, Alvis Brazma1, Zemin Zhang2 1 European Molecular Biology Laboratory, European Bioinformatics Institute (EMBL-EBI), Wellcome Trust Genome Campus, Hinxton, Cambridge CB10 1SD, UK; 2 Peking-Tsinghua Centre for Life Sciences, BIOPIC, and Beijing Advanced Innovation Centre for Genomics, Peking University, Beijing, 100871, China *Joint first authors Gene fusions are an important class of cancer-driving events with therapeutic and diagnostic values, yet their underlying genetic mechanisms have not been systematically characterized. Here by combining RNA and whole genome DNA sequencing data from 1188 donors across 27 cancer types we obtained a list of 3297 high-confidence tumour-specific gene fusions, 82% of which had structural variant (SV) support and 2372 of which were novel. Such a large collection of RNA and DNA alterations provides the first opportunity to systematically classify the gene fusions at a mechanistic level. While many could be explained by single SVs, numerous fusions involved series of structural rearrangements and thus are composite fusions. We discovered 75 fusions of a novel class of inter-chromosomal composite fusions, termed bridged fusions, in which a third genomic location bridged two different genes. In addition, we identified 522 fusions involving non-coding genes and 157 ORF-retaining fusions, in which the complete open reading frame of one gene was fused to the UTR region of another. Although only a small proportion (5%) of the discovered fusions were recurrent, we found a set of highly recurrent fusion partner genes, which exhibited strong 5’ or 3’ bias and were significantly enriched for cancer genes. Our findings broaden the view of the gene fusion landscape and reveal the general properties of genetic alterations underlying gene fusions for the first time. The simultaneous availability of whole genome and RNA sequencing data for a large panel of International Cancer Genome Consortium (ICGC) tumour samples creates an unprecedented opportunity to identify novel gene fusions and understand their genetic basis. Gene fusions were detected by combining the output of two fusion discovery methods and genomic rearrangement (structural variant - SV) information, and several filters were implemented to exclude artefacts or those also present in the ICGC or GTEx normal samples [1] (Online Methods; Supplementary Fig 1; Supplementary Fig 2). Gene fusions were associated to the nearest SV with both breakpoints located within 250 kb of the two corresponding fusion breakpoints. These gene fusions were categorized based on novelty, recurrence, known 1 bioRxiv preprint doi: https://doi.org/10.1101/148684; this version posted June 12, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. oncogenic gene partners, breakpoint distance, breakpoint location, and matched SV support for downstream analyses (Fig. 1). The average number of putative gene fusions per sample varies considerably across histological types (mean=3, median=2, sd=3) and the number of gene fusions is highly correlated with the average number of SVs (Pearson correlation 0.96), supporting SVs as a major cause of gene fusions (Fig. 1B). For instance, bone- leimyocarcinomas harbour ~14 fusions per sample on average, while 10 cancer types have less than one fusion per sample. The fusion frequencies were also associated with other types of alterations (Supplementary Fig. 3) but the observed correlation is lower (Supplementary Fig. 4). By examining somatic rearrangement events and fusions simultaneously, we found 2618 fusion events that could be explained by single genomic rearrangements, with duplication as the predominant type, suggesting genomic amplification as a major source for gene fusions (Supplementary Fig. 5). Although 1288 of these rearrangements occurred within gene bodies by juxtaposing exons from two different genes, 720 other rearrangements disrupted neither of the fusion gene partners (Supplementary Fig. 5). Notably, a large number of fusions, including known fusions (Supplementary Fig. 6) namely ETV6-NTRK3 [2] could not be associated with any single SV event. The ETV6-NTRK3 fusion was present in a head and neck thyroid carcinoma sample, linking exon 4 of ETV6 to exon 12 of NTRK3. We found three separate SVs: i) a translocation of ETV6 (chr12:12,099,706) to chromosome 6 (chr6:125,106,892); ii) a translocation of NTRK3 (chr15:88,694,049) also to chromosome 6 (chr6:125,062,387) (Fig. 2B); and iii) an additional copy number loss (chr12:12,032,501-chr12:12,099,705) spanning from ETV6 intron 5 to the exact SV breakpoints (chr12:12,099,706), jointly bringing ETV6 within 45 kb upstream of NTRK3, a distance that would allow transcriptional read-through [3] or splicing [4] to yield the ETV6-NTRK3 fusion [5]. Thus, the short chromosome 6 segment appeared to function as a bridge, linking two other genomic locations to facilitate a gene fusion. We term such products bridged fusions. This novel class of fusions are not uncommon. Out of a total of 436 fusions supported by two separate SVs, 75 are bridged fusions. The lengths of bridge regions ranges from 1bp to 496 kb, with a median size of 3.7 kb and an average size of 62.96 kb (Supplementary Fig. 7). One third of the bridge regions are located in genic regions and none of such genic regions were found to be involved in other fusions in the corresponding samples. Aside from bridged fusions, 344 additional fusions are supported by more than one SV. These multi-SV fusions are collectively termed composite fusions. For example, the known ERC1-RET fusion, previously described as a product of the translocation t(10;12)(q11;p13), was detected in two samples of head and neck and thyroid carcinomas. While there was no evidence of direct translocation in either case, they were both supported by an inter-chromosomal translocation and an intra- chromosomal rearrangement (deletion or inversion), resulting in the connection of either exon 12 or exon 17 of ERC1 to the exon 12 of RET (Fig. 2C). 2 bioRxiv preprint doi: https://doi.org/10.1101/148684; this version posted June 12, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. While fusion transcripts formed by two adjacent genes are often viewed as non- genomic transcription-induced chimeras [3], we observed that such chimera formation could possibly be facilitated by composite DNA rearrangements. For one of the tumours with the recurrent NUMB-HEATR4 fusion, we detected two consecutive inversions, bringing the NUMB exon 3 within 381 bp of the HEATR4 exon 2 (Fig. 2D). The much shorter distance, down from the natural distance of 14 kb, would allow fusion formation by splicing. Such fusion could also result from duplication – it was supported by a duplication spanning from exon 2 of the NUMB gene to exon 4 of the HEATR4 gene in two other tumours. Based on the nature of underlying genomic rearrangements, we propose a unified fusion classification system (Fig. 2E). Overall, we identified 75 bridged fusions, 284 inter-composite fusions generated by a translocation linking two genes from different chromosomes followed by a second intra-chromosomal rearrangement, and 125 intra-composite fusions generated by multiple intra-chromosomal rearrangements. To further understand the mechanisms of composite fusion formation, we examined various features associated with such fusions. Intra-composite fusion partners were brought significantly closer to each other, from the median natural distance of 6,836 kb to the median of 7.9 kb (Wilcoxon Rank Sum Test, P < 2.2e-16, Fig. 3A). Inter- composite fusion partners also exhibited similarly short gene distances post- translocation (Fig. 3A). Although bridged fusion pillar breakpoints were distributed broadly across the genome with no particular hotspots (Supplementary Fig. 8), they were over-represented on chromosome 12 (Fig. 3C, Supplementary Fig. 9). Notably, such pillar breakpoints were associated with genomic regions with high mutation density or high rearrangement density (Fig. 3C), indicating the connection with mutational processes such as chromothripsis and kataegis [6,7] (Fig. 3C). Eighty eight of the 150 pillar breakpoints overlapped with chromothripsis regions [Peter Park personal communication], and 33 of these 88 regions further accompanied kataegis [8]. For the remaining 62 pillar breakpoints, while we did not find local enrichment for any specific sequence motifs or fragile sites, 53% of these were located within the Alu and LINE repeat elements or other simple repeats. Based on the presence of microhomology in the breakpoints and the hypermutation status of bridge regions, we inferred that 23 of the 75 bridged fusions were formed by replication-based DNA double strand repair process, and 52 fusions were formed by either non-homologous or microhomology-mediated end-joining (NHEJ/MMEJ) (Fig. 3D). Thus, multiple mechanisms could lead to bridged fusions, with chromothripsis playing a key role. While most fusions had direct or composite SV support, the remaining 18%, including known fusions like RHOH-BCL6 [9] did not have obvious SV evidence. Thus, either these genes were fused directly at the RNA level or the underlying supporting SVs were somehow missed.